Graphics processing units (GPUs) may contain relatively large and dense caches to support highly parallelized execution. The density and size of the GPU caches may make them vulnerable to failure, particularly over time (e.g., after deployment in the field). While error correction code (ECC) solutions may address GPU cache failure concerns to some extent, there remains considerable room for improvement.